Appendix

Theexpected performance objective in (13) is equivalent setting toβ = 0 in Figure 1a. The algorithm gets stuck and repeatedly evaluates inputs (orange crosses) at alocal optimum of the true objectivefunction (black dashed).